Multi-Task Gravity and Magnetic Inversion Based on Deep Learning

Gravity and magnetic inversions are the processes of utilizing anomaly data collected at the surface to obtain the subsurface spatial structure and physical properties. In recent years, deep learning- based gravity and magnetic inversion methods have achieved good results. However, the inherent non-uniqueness of the inversion problem makes the inversion results unfocused and perform poorly on multiple density values. Therefore, we propose to use multi-task learning to learn multiple related tasks simultaneously to improve the network generalization ability. Here, a multi-task UNet3+ network is used to achieve both anomaly localization and density value reconstruction. Joint inversion is also considered to synthesize relevant geophysical data with different physical properties in the same survey area. Different data complement each other to enhance the credibility of the inversion results and reduce the non-uniqueness. The experiments show that the multi-task method is more focused and accurate than the single-task method, and the multi-source data can further improve the inversion effect in the depth.